01 8× NVIDIA H1000/RTX5090 SXM · CONFIDENTIAL COMPUTE
02 SEALED CHASSIS · TAMPER-EVIDENT · GOV-ISSUED KEYS
Sovereign AI // v3 VW-SAI-DC-08 Deployable in weeks

The GPU sees
computation.
It never sees data.

VEILWATCH Sovereign AI is a purpose-built hardware and software stack for sovereign-nation intelligence work — biometric matching, document authentication, video analytics, and risk inference. Models are trained, served, and audited entirely inside the government perimeter. Compute scales out without ever exposing what is being computed.

GPU CAPACITY
2.4 PFLOPS · FP8
BIOMETRIC 1:M
<420 MS @ 50M
DOC CORPUS
193 STATES · 7 LAYERS
SOVEREIGN
100% WEIGHTS · DATA
SECTION 01OPERATING PRINCIPLE

Tier-1 compute power.
Air-gapped data sovereignty.

Reference:
13.3 Distributed GPU Compute
13.4 Sovereign Training

VEILWATCH Sovereign AI separates what runs from what is run on. Models, weights, and orchestration sit inside the government perimeter. Raw biometric, identity, and intelligence data never leaves.

Sensitive data is mathematically transformed into non-invertible inference payloads before any compute is dispatched. GPU nodes — sovereign-owned, on-premises, or hybrid — receive computation tasks with no access to the underlying personal, biometric, or intelligence data. Aggregated inference results are returned and re-associated with subject records inside the sovereign perimeter.

Hardware-level confidential compute — Intel TDX, AMD SEV-SNP, and NVIDIA H1000/RTX5090 Confidential Computing — provides cryptographic attestation that even the GPU operator cannot read what is being computed. Every dispatch and inference return is logged with cryptographic proof of execution.

VW.PRINCIPLE // ZK-DISPATCH

Raw data does not cross the perimeter. Only computation does.

VW-SAI ARCHITECTURE PRINCIPLE 001 · 13.3 / 13.4
SECTION 02STACK ARCHITECTURE

Three layers,
one perimeter.

Component Map:
SOV PERIMETER // ZK LAYER // COMPUTE
VW-SAI // ARCHITECTURE DIAGRAM · FIG. 02

Zero-knowledge GPU dispatch, end to end.

A request enters the sovereign perimeter from a border post, port, or operations centre. Data is transformed inside the perimeter. Only the transformed payload — never raw data — crosses into the compute fabric.

L01 // TRUSTED

Sovereign Perimeter

Inside // Government custody

Where raw data lives, is captured, and is destroyed. Holds the only mapping between subject identity and inference results.

  • Capture endpoints
    Tablets, kiosks, cameras, NFC
  • Identity vault
    Encrypted at rest · HSM-bound
  • Result re-association
    Inference ↔ subject mapping
  • Audit ledger
    Tamper-evident · WORM
L02 // TRANSFORM

Zero-Knowledge Layer

Boundary // Cryptographic

The perimeter wall. Transforms raw data into non-invertible inference payloads, attests the destination GPU, and signs every dispatch.

  • Payload transform
    Non-invertible & sharded
  • Remote attestation
    TDX · SEV-SNP · H1000/RTX5090 CC
  • Model sharding
    No node holds full model
  • Dispatch proof
    Cryptographically logged
L03 // OPAQUE

Compute Fabric

Outside // Sees math only

Sovereign GPU cluster, government on-premises, or vetted external bursting. Sees inference math; never sees subjects, biometrics, or identifiers.

  • Sovereign fleet
    VW-owned · physically isolated
  • On-prem cluster
    Customer hardware · air-gapped
  • Hybrid burst
    Vetted external · ZK-only
  • Encrypted fabric
    Logical & physical isolation
CAPTURE
Raw subject data
TRANSFORM
L02 · ZK DISPATCH
Non-invertible payload
INFER
RETURN · L01
Aggregated result
SIGNAL — TRANSFORMED PAYLOAD
VERIFIED — ATTESTED RETURN
BLOCKED — RAW EGRESS
PERIMETER — SOVEREIGN BOUNDARY
SECTION 03HARDWARE PLATFORM

VW-SAI Rack // Sovereign appliance.

Form Factor:
42U · TAMPER-EVIDENT
SEALED CHASSIS
VW-SAI / RACK · SKU 0801
CHASSIS
42U / SEALED
DRAW
22.4 kW
COOL
DLC // N+1
VW-SAI Rack · 42U Sovereign Appliance
SKU // VW-SAI-0801
U41–42
Top-of-Rack Switching
2× 800 GbE · IPsec offload · MACsec
×2
U33–40
GPU Compute — VEILWATCH H1000/RTX5090 Node
8× NVIDIA H1000/RTX5090 SXM5 · NVLink · Confidential Compute
×4
U25–32
Inference Node — VW-INF-A
2× Intel Xeon TDX · 2 TB DDR5 · 4× L40S
×4
U17–24
HSM & Attestation Cluster
FIPS 140-3 L4 · Quorum-bound key release
×2
U09–16
Sovereign Storage Array
1.2 PB NVMe · LUKS2 + per-tenant XTS · WORM ledger
×1
U05–08
Network Boundary & Diode
One-way data diode · DPI · Out-of-band mgmt
×1
U03–04
Direct Liquid Cooling Manifold
Rear-door HX · N+1 pumps · leak detection
×1
U01–02
PDU — Dual-Feed
2× 32 A · monitored · sealed breaker bay
A · B
FP8 PEAK
2.4 PF
VRAM
2.56 TB
DRAW
22.4 kW
WEIGHT
1,180 kg

Built to deploy.
Built to be trusted.

Sealed chassis, government-issued keys, factory-attested firmware, dual-feed power, and an out-of-band management plane that cannot be reached from operational networks. Configured, sealed, and shipped from VEILWATCH NA. Receipt-verified at port. Operational from day one.

DEPLOY TIME
14 DAYS · TYPICAL
From dock to first inference. No lengthy procurement cycle, no greenfield datacenter dependency.
RAW DATA EGRESS
BYTES
Contractually guaranteed. Cryptographically attested. Continuously audited.
SECTION 04AI MODEL STACK

Six models.
All weights sovereign.

Lifecycle:
VW MODEL CONTROL
SIGNED // VERSIONED
VW-FRX // 01 v6.4.2 · 2026Q1

Facial Geometry NN

arch · ResNeXt-Vision · 1024-d landmark embedding

Deep-learning facial recognition trained on billions of landmark configurations. Sub-second 1:1 and 1:M biometric matching at an error rate two orders of magnitude lower than human review.

1:M @ 50M
412 MS
FMR
1e-7
VW-LIV // 02 v4.1.0 · 2026Q1

Liveness — Anti-Spoof

arch · Multi-frame transformer · passive + active

Distinguishes live presentations from printed photos, screen replay, silicone masks, and deepfake injections. Passive mode confirms liveness without operator interaction.

PAD APCER
0.04 %
DEEPFAKE
99.6 % TPR
VW-DOC // 03 v8.0.1 · 2026Q1

Document Authentication

arch · 7-layer ensemble · per-issuing-state heads

Layered authentication of MRZ structure, visual zone, NFC chip, biographic consistency, issuing-state pattern, security features, and historical cross-reference. Each fail emits a fraud indicator code.

STATES
193
LAYERS
7
VW-MRZ // 04 v5.7.3 · 2026Q1

MRZ Vision Engine

arch · Hybrid CNN · field-conditioned OCR

Proprietary MRZ extraction trained on degraded, worn, photocopied, and damaged documents — conditions where commercial OCR libraries fail. Sub-500 ms extraction from a live camera feed.

EXTRACT
<500 MS
FORMATS
TD1/2/3
VW-VID // 05 v3.2.0 · 2026Q1

Neural Video Analytics

arch · Multi-stream YOLO-derivative · re-ID head

Transforms passive CCTV into active intelligence. Real-time person, vehicle, and behaviour detection across hundreds of streams; persistent re-identification across cameras and ports.

STREAMS
512 / NODE
FPS
30
VW-RSK // 06 v7.1.4 · 2026Q1

SENTINEL Risk Engine

arch · Gradient-boosted ensemble · graph features

AI-driven risk scoring and dynamic threat profiling. Fuses NEXUS background intelligence, ISR vessel/air data, and BORDER processing signals. Every score carries an explainability payload.

SIGNALS
320+
EXPLAINABLE
100 %
SECTION 05DEPLOYMENT MODELS

Three deployment paths.
One trust model.

All Modes:
ZK-DISPATCH ENFORCED
MODEL WEIGHTS ENCRYPTED
MODE 01 // PRIMARY RECOMMENDED

Sovereign GPU Fleet

VEILWATCH operates a dedicated, physically isolated GPU cluster for the customer government. No shared tenancy. No co-located workloads. Direct fibre to the sovereign perimeter.

  • OWNERSHIPVEILWATCH-managed, customer-dedicated
  • LOCATIONVW-NA / VW-EU sovereign datacenters
  • TENANCYSingle-tenant, physically isolated
  • BURSTVertical scaling within fleet
  • BEST FORNational border & immigration platforms
MODE 02 // ON-PREM AIR-GAPPED

Government Datacenter

Full GPU cluster deployed inside the customer's own datacenter. VEILWATCH models and orchestration layer run on customer-owned hardware with no external connectivity required.

  • OWNERSHIPCustomer-owned hardware
  • LOCATIONNational datacenter / military facility
  • TENANCYAir-gapped, no internet egress
  • BURSTOptional ZK burst to sovereign fleet
  • BEST FORDefence, intelligence, regulated finance
MODE 03 // HYBRID ELASTIC

ZK External Burst

Baseline runs sovereign or on-premises. Overflow inference bursts to vetted external GPU providers using the zero-knowledge dispatch layer — elastic capacity without sovereign compromise.

  • OWNERSHIPHybrid · vetted external compute
  • LOCATIONCustomer-approved provider list
  • TENANCYMulti-tenant, ZK-only payloads
  • BURSTElastic — peak event scaling
  • BEST FORMajor events, peak season, surge
SECTION 06SECURITY & COMPLIANCE

Verifiable, not
just stated.

Independent:
CERTIFIED // TESTED
CONTRACTUALLY BOUND

Every model and every dispatch
is tested, signed, and audited.

VEILWATCH AI models deployed in government border operations are independently tested and certified prior to production deployment. Performance is contractually specified. Bias is continuously measured across demographic groups, document origins, and environmental conditions.

FIPS 140-3 HSM & key releaseLevel 4 cryptographic module · quorum-bound
ICAO 9303 Travel document authenticationTD1, TD2, TD3 across all issuing states
ISO 18013-5 Mobile driving licenceNFC + QR verification compliant
ISO 30107 Presentation attack detectionPAD compliance · level 2 tested
SOC 2 TYPE II Operational controlsAnnual independent attestation
NIST FRVT Face recognition vendor test1:1 and 1:N benchmarks · published results
CJIS Law-enforcement data handlingCompatible deployment profile
USPTO Sensor fusion methodologyProvisional application filed (VEILWATCH Coastal)